Evidence boundary

The workshop originates in analysis run AAS-2026-08-27-commonplace-review-freshness-01. The run froze revision c5afbf1c5efff02d47d71c4bf05779fa5170ab4a and inspected the review/freshness subsystem from selection through result read-back. At workshop creation, the registered subsystem files remained byte-equivalent at checkout revision 64e84517712ffa61682bd3f63ba859223f9330cc.

Inspected material

Observed checks

The complete test suite passed: 580 passed in 30.01s. Disposable-repository interventions also established that:

  • finalizing FAIL creates a fresh baseline, removes the pair from stale selection, and does not place it in the warning queue;
  • changing only linked evidence leaves a completed pair fresh;
  • changing only derived result prose changes warning read-back while the SQLite outcome stays unchanged;
  • an invalid declared snapshot hash and canonical source do not prevent model-shaped PASS output from finalizing and becoming fresh;
  • failure in a later creation group leaves earlier queued state and the failing job in SQLite without returning their normal JSON handles; and
  • extract_warns() returns an explicit - warn: item from PASS-shaped input.

These interventions establish deterministic subsystem behavior. They do not establish what a properly instructed external reviewer would read, infer, or return.

Excluded evidence

No real provider trace, parent-harness dispatch trace, production review store, operator disposition history, or gate-calibration corpus was inspected. The workshop therefore treats activation frequency, reviewer accuracy, production incidence, and model-equivalence quality as unresolved.